err.no Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/pci.h>
  35 #include <linux/poll.h>
  36 #include <linux/cdev.h>
  37 #include <linux/swap.h>
  38 #include <linux/vmalloc.h>
  39 #include <asm/pgtable.h>
  40
  41 #include "ipath_kernel.h"
  42 #include "ips_common.h"
  43 #include "ipath_layer.h"
  44
  45 static int ipath_open(struct inode *, struct file *);
  46 static int ipath_close(struct inode *, struct file *);
  47 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  48                            loff_t *);
  49 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  50 static int ipath_mmap(struct file *, struct vm_area_struct *);
  51
  52 static struct file_operations ipath_file_ops = {
  53         .owner = THIS_MODULE,
  54         .write = ipath_write,
  55         .open = ipath_open,
  56         .release = ipath_close,
  57         .poll = ipath_poll,
  58         .mmap = ipath_mmap
  59 };
  60
  61 static int ipath_get_base_info(struct ipath_portdata *pd,
  62                                void __user *ubase, size_t ubase_size)
  63 {
  64         int ret = 0;
  65         struct ipath_base_info *kinfo = NULL;
  66         struct ipath_devdata *dd = pd->port_dd;
  67
  68         if (ubase_size < sizeof(*kinfo)) {
  69                 ipath_cdbg(PROC,
  70                            "Base size %lu, need %lu (version mismatch?)\n",
  71                            (unsigned long) ubase_size,
  72                            (unsigned long) sizeof(*kinfo));
  73                 ret = -EINVAL;
  74                 goto bail;
  75         }
  76
  77         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
  78         if (kinfo == NULL) {
  79                 ret = -ENOMEM;
  80                 goto bail;
  81         }
  82
  83         ret = dd->ipath_f_get_base_info(pd, kinfo);
  84         if (ret < 0)
  85                 goto bail;
  86
  87         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
  88         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
  89         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
  90         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
  91         /*
  92          * have to mmap whole thing
  93          */
  94         kinfo->spi_rcv_egrbuftotlen =
  95                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
  96         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
  97         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
  98                 pd->port_rcvegrbuf_chunks;
  99         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt;
 100         /*
 101          * for this use, may be ipath_cfgports summed over all chips that
 102          * are are configured and present
 103          */
 104         kinfo->spi_nports = dd->ipath_cfgports;
 105         /* unit (chip/board) our port is on */
 106         kinfo->spi_unit = dd->ipath_unit;
 107         /* for now, only a single page */
 108         kinfo->spi_tid_maxsize = PAGE_SIZE;
 109
 110         /*
 111          * Doing this per port, and based on the skip value, etc.  This has
 112          * to be the actual buffer size, since the protocol code treats it
 113          * as an array.
 114          *
 115          * These have to be set to user addresses in the user code via mmap.
 116          * These values are used on return to user code for the mmap target
 117          * addresses only.  For 32 bit, same 44 bit address problem, so use
 118          * the physical address, not virtual.  Before 2.6.11, using the
 119          * page_address() macro worked, but in 2.6.11, even that returns the
 120          * full 64 bit address (upper bits all 1's).  So far, using the
 121          * physical addresses (or chip offsets, for chip mapping) works, but
 122          * no doubt some future kernel release will chang that, and we'll be
 123          * on to yet another method of dealing with this
 124          */
 125         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 126         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 127         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 128         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 129                 (void *) dd->ipath_statusp -
 130                 (void *) dd->ipath_pioavailregs_dma;
 131         kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 132         kinfo->__spi_uregbase =
 133                 dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
 134
 135         kinfo->spi_pioindex = dd->ipath_pbufsport * (pd->port_port - 1);
 136         kinfo->spi_piocnt = dd->ipath_pbufsport;
 137         kinfo->spi_pioalign = dd->ipath_palign;
 138
 139         kinfo->spi_qpair = IPATH_KD_QP;
 140         kinfo->spi_piosize = dd->ipath_ibmaxlen;
 141         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 142         kinfo->spi_port = pd->port_port;
 143         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 144         kinfo->spi_hw_version = dd->ipath_revision;
 145
 146         if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
 147                 ret = -EFAULT;
 148
 149 bail:
 150         kfree(kinfo);
 151         return ret;
 152 }
 153
 154 /**
 155  * ipath_tid_update - update a port TID
 156  * @pd: the port
 157  * @ti: the TID information
 158  *
 159  * The new implementation as of Oct 2004 is that the driver assigns
 160  * the tid and returns it to the caller.   To make it easier to
 161  * catch bugs, and to reduce search time, we keep a cursor for
 162  * each port, walking the shadow tid array to find one that's not
 163  * in use.
 164  *
 165  * For now, if we can't allocate the full list, we fail, although
 166  * in the long run, we'll allocate as many as we can, and the
 167  * caller will deal with that by trying the remaining pages later.
 168  * That means that when we fail, we have to mark the tids as not in
 169  * use again, in our shadow copy.
 170  *
 171  * It's up to the caller to free the tids when they are done.
 172  * We'll unlock the pages as they free them.
 173  *
 174  * Also, right now we are locking one page at a time, but since
 175  * the intended use of this routine is for a single group of
 176  * virtually contiguous pages, that should change to improve
 177  * performance.
 178  */
 179 static int ipath_tid_update(struct ipath_portdata *pd,
 180                             const struct ipath_tid_info *ti)
 181 {
 182         int ret = 0, ntids;
 183         u32 tid, porttid, cnt, i, tidcnt;
 184         u16 *tidlist;
 185         struct ipath_devdata *dd = pd->port_dd;
 186         u64 physaddr;
 187         unsigned long vaddr;
 188         u64 __iomem *tidbase;
 189         unsigned long tidmap[8];
 190         struct page **pagep = NULL;
 191
 192         if (!dd->ipath_pageshadow) {
 193                 ret = -ENOMEM;
 194                 goto done;
 195         }
 196
 197         cnt = ti->tidcnt;
 198         if (!cnt) {
 199                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 200                           (unsigned long long) ti->tidlist);
 201                 /*
 202                  * Should we treat as success?  likely a bug
 203                  */
 204                 ret = -EFAULT;
 205                 goto done;
 206         }
 207         tidcnt = dd->ipath_rcvtidcnt;
 208         if (cnt >= tidcnt) {
 209                 /* make sure it all fits in port_tid_pg_list */
 210                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 211                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 212                 cnt = tidcnt;
 213         }
 214         pagep = (struct page **)pd->port_tid_pg_list;
 215         tidlist = (u16 *) (&pagep[cnt]);
 216
 217         memset(tidmap, 0, sizeof(tidmap));
 218         tid = pd->port_tidcursor;
 219         /* before decrement; chip actual # */
 220         porttid = pd->port_port * tidcnt;
 221         ntids = tidcnt;
 222         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 223                                    dd->ipath_rcvtidbase +
 224                                    porttid * sizeof(*tidbase));
 225
 226         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 227                    pd->port_port, cnt, tid, tidbase);
 228
 229         /* virtual address of first page in transfer */
 230         vaddr = ti->tidvaddr;
 231         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 232                        cnt * PAGE_SIZE)) {
 233                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 234                           (void *)vaddr, cnt);
 235                 ret = -EFAULT;
 236                 goto done;
 237         }
 238         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 239         if (ret) {
 240                 if (ret == -EBUSY) {
 241                         ipath_dbg("Failed to lock addr %p, %u pages "
 242                                   "(already locked)\n",
 243                                   (void *) vaddr, cnt);
 244                         /*
 245                          * for now, continue, and see what happens but with
 246                          * the new implementation, this should never happen,
 247                          * unless perhaps the user has mpin'ed the pages
 248                          * themselves (something we need to test)
 249                          */
 250                         ret = 0;
 251                 } else {
 252                         dev_info(&dd->pcidev->dev,
 253                                  "Failed to lock addr %p, %u pages: "
 254                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 255                         goto done;
 256                 }
 257         }
 258         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 259                 for (; ntids--; tid++) {
 260                         if (tid == tidcnt)
 261                                 tid = 0;
 262                         if (!dd->ipath_pageshadow[porttid + tid])
 263                                 break;
 264                 }
 265                 if (ntids < 0) {
 266                         /*
 267                          * oops, wrapped all the way through their TIDs,
 268                          * and didn't have enough free; see comments at
 269                          * start of routine
 270                          */
 271                         ipath_dbg("Not enough free TIDs for %u pages "
 272                                   "(index %d), failing\n", cnt, i);
 273                         i--;    /* last tidlist[i] not filled in */
 274                         ret = -ENOMEM;
 275                         break;
 276                 }
 277                 tidlist[i] = tid;
 278                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 279                            "vaddr %lx\n", i, tid, vaddr);
 280                 /* we "know" system pages and TID pages are same size */
 281                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 282                 /*
 283                  * don't need atomic or it's overhead
 284                  */
 285                 __set_bit(tid, tidmap);
 286                 physaddr = page_to_phys(pagep[i]);
 287                 ipath_stats.sps_pagelocks++;
 288                 ipath_cdbg(VERBOSE,
 289                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 290                            tid, vaddr, (unsigned long long) physaddr,
 291                            pagep[i]);
 292                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
 293                 /*
 294                  * don't check this tid in ipath_portshadow, since we
 295                  * just filled it in; start with the next one.
 296                  */
 297                 tid++;
 298         }
 299
 300         if (ret) {
 301                 u32 limit;
 302         cleanup:
 303                 /* jump here if copy out of updated info failed... */
 304                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 305                           -ret, i, cnt);
 306                 /* same code that's in ipath_free_tid() */
 307                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 308                 if (limit > tidcnt)
 309                         /* just in case size changes in future */
 310                         limit = tidcnt;
 311                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 312                 for (; tid < limit; tid++) {
 313                         if (!test_bit(tid, tidmap))
 314                                 continue;
 315                         if (dd->ipath_pageshadow[porttid + tid]) {
 316                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 317                                            tid);
 318                                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
 319                                                     dd->ipath_tidinvalid);
 320                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 321                                 ipath_stats.sps_pageunlocks++;
 322                         }
 323                 }
 324                 ipath_release_user_pages(pagep, cnt);
 325         } else {
 326                 /*
 327                  * Copy the updated array, with ipath_tid's filled in, back
 328                  * to user.  Since we did the copy in already, this "should
 329                  * never fail" If it does, we have to clean up...
 330                  */
 331                 if (copy_to_user((void __user *)
 332                                  (unsigned long) ti->tidlist,
 333                                  tidlist, cnt * sizeof(*tidlist))) {
 334                         ret = -EFAULT;
 335                         goto cleanup;
 336                 }
 337                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 338                                  tidmap, sizeof tidmap)) {
 339                         ret = -EFAULT;
 340                         goto cleanup;
 341                 }
 342                 if (tid == tidcnt)
 343                         tid = 0;
 344                 pd->port_tidcursor = tid;
 345         }
 346
 347 done:
 348         if (ret)
 349                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 350                           ti->tidcnt, -ret);
 351         return ret;
 352 }
 353
 354 /**
 355  * ipath_tid_free - free a port TID
 356  * @pd: the port
 357  * @ti: the TID info
 358  *
 359  * right now we are unlocking one page at a time, but since
 360  * the intended use of this routine is for a single group of
 361  * virtually contiguous pages, that should change to improve
 362  * performance.  We check that the TID is in range for this port
 363  * but otherwise don't check validity; if user has an error and
 364  * frees the wrong tid, it's only their own data that can thereby
 365  * be corrupted.  We do check that the TID was in use, for sanity
 366  * We always use our idea of the saved address, not the address that
 367  * they pass in to us.
 368  */
 369
 370 static int ipath_tid_free(struct ipath_portdata *pd,
 371                           const struct ipath_tid_info *ti)
 372 {
 373         int ret = 0;
 374         u32 tid, porttid, cnt, limit, tidcnt;
 375         struct ipath_devdata *dd = pd->port_dd;
 376         u64 __iomem *tidbase;
 377         unsigned long tidmap[8];
 378
 379         if (!dd->ipath_pageshadow) {
 380                 ret = -ENOMEM;
 381                 goto done;
 382         }
 383
 384         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 385                            sizeof tidmap)) {
 386                 ret = -EFAULT;
 387                 goto done;
 388         }
 389
 390         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 391         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 392                                    dd->ipath_rcvtidbase +
 393                                    porttid * sizeof(*tidbase));
 394
 395         tidcnt = dd->ipath_rcvtidcnt;
 396         limit = sizeof(tidmap) * BITS_PER_BYTE;
 397         if (limit > tidcnt)
 398                 /* just in case size changes in future */
 399                 limit = tidcnt;
 400         tid = find_first_bit(tidmap, limit);
 401         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 402                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 403                    limit, tid, porttid);
 404         for (cnt = 0; tid < limit; tid++) {
 405                 /*
 406                  * small optimization; if we detect a run of 3 or so without
 407                  * any set, use find_first_bit again.  That's mainly to
 408                  * accelerate the case where we wrapped, so we have some at
 409                  * the beginning, and some at the end, and a big gap
 410                  * in the middle.
 411                  */
 412                 if (!test_bit(tid, tidmap))
 413                         continue;
 414                 cnt++;
 415                 if (dd->ipath_pageshadow[porttid + tid]) {
 416                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 417                                    pd->port_pid, tid);
 418                         dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
 419                                             dd->ipath_tidinvalid);
 420                         ipath_release_user_pages(
 421                                 &dd->ipath_pageshadow[porttid + tid], 1);
 422                         dd->ipath_pageshadow[porttid + tid] = NULL;
 423                         ipath_stats.sps_pageunlocks++;
 424                 } else
 425                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 426         }
 427         if (cnt != ti->tidcnt)
 428                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 429                           ti->tidcnt, cnt);
 430 done:
 431         if (ret)
 432                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 433                           ti->tidcnt, -ret);
 434         return ret;
 435 }
 436
 437 /**
 438  * ipath_set_part_key - set a partition key
 439  * @pd: the port
 440  * @key: the key
 441  *
 442  * We can have up to 4 active at a time (other than the default, which is
 443  * always allowed).  This is somewhat tricky, since multiple ports may set
 444  * the same key, so we reference count them, and clean up at exit.  All 4
 445  * partition keys are packed into a single infinipath register.  It's an
 446  * error for a process to set the same pkey multiple times.  We provide no
 447  * mechanism to de-allocate a pkey at this time, we may eventually need to
 448  * do that.  I've used the atomic operations, and no locking, and only make
 449  * a single pass through what's available.  This should be more than
 450  * adequate for some time. I'll think about spinlocks or the like if and as
 451  * it's necessary.
 452  */
 453 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 454 {
 455         struct ipath_devdata *dd = pd->port_dd;
 456         int i, any = 0, pidx = -1;
 457         u16 lkey = key & 0x7FFF;
 458         int ret;
 459
 460         if (lkey == (IPS_DEFAULT_P_KEY & 0x7FFF)) {
 461                 /* nothing to do; this key always valid */
 462                 ret = 0;
 463                 goto bail;
 464         }
 465
 466         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 467                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 468                    pd->port_port, key, dd->ipath_pkeys[0],
 469                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 470                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 471                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 472                    atomic_read(&dd->ipath_pkeyrefs[3]));
 473
 474         if (!lkey) {
 475                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 476                            pd->port_port);
 477                 ret = -EINVAL;
 478                 goto bail;
 479         }
 480
 481         /*
 482          * Set the full membership bit, because it has to be
 483          * set in the register or the packet, and it seems
 484          * cleaner to set in the register than to force all
 485          * callers to set it. (see bug 4331)
 486          */
 487         key |= 0x8000;
 488
 489         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 490                 if (!pd->port_pkeys[i] && pidx == -1)
 491                         pidx = i;
 492                 if (pd->port_pkeys[i] == key) {
 493                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 494                                    "(%x) more than once\n",
 495                                    pd->port_port, key);
 496                         ret = -EEXIST;
 497                         goto bail;
 498                 }
 499         }
 500         if (pidx == -1) {
 501                 ipath_dbg("All pkeys for port %u already in use, "
 502                           "can't set %x\n", pd->port_port, key);
 503                 ret = -EBUSY;
 504                 goto bail;
 505         }
 506         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 507                 if (!dd->ipath_pkeys[i]) {
 508                         any++;
 509                         continue;
 510                 }
 511                 if (dd->ipath_pkeys[i] == key) {
 512                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 513
 514                         if (atomic_inc_return(pkrefs) > 1) {
 515                                 pd->port_pkeys[pidx] = key;
 516                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 517                                            "matches #%d, count now %d\n",
 518                                            pd->port_port, key, i,
 519                                            atomic_read(pkrefs));
 520                                 ret = 0;
 521                                 goto bail;
 522                         } else {
 523                                 /*
 524                                  * lost race, decrement count, catch below
 525                                  */
 526                                 atomic_dec(pkrefs);
 527                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 528                                            "0, after dec, it's %d\n",
 529                                            atomic_read(pkrefs));
 530                                 any++;
 531                         }
 532                 }
 533                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 534                         /*
 535                          * It makes no sense to have both the limited and
 536                          * full membership PKEY set at the same time since
 537                          * the unlimited one will disable the limited one.
 538                          */
 539                         ret = -EEXIST;
 540                         goto bail;
 541                 }
 542         }
 543         if (!any) {
 544                 ipath_dbg("port %u, all pkeys already in use, "
 545                           "can't set %x\n", pd->port_port, key);
 546                 ret = -EBUSY;
 547                 goto bail;
 548         }
 549         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 550                 if (!dd->ipath_pkeys[i] &&
 551                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 552                         u64 pkey;
 553
 554                         /* for ipathstats, etc. */
 555                         ipath_stats.sps_pkeys[i] = lkey;
 556                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 557                         pkey =
 558                                 (u64) dd->ipath_pkeys[0] |
 559                                 ((u64) dd->ipath_pkeys[1] << 16) |
 560                                 ((u64) dd->ipath_pkeys[2] << 32) |
 561                                 ((u64) dd->ipath_pkeys[3] << 48);
 562                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 563                                    "portidx %d, new pkey reg %llx\n",
 564                                    pd->port_port, key, i, pidx,
 565                                    (unsigned long long) pkey);
 566                         ipath_write_kreg(
 567                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 568
 569                         ret = 0;
 570                         goto bail;
 571                 }
 572         }
 573         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 574                   "can't set %x\n", pd->port_port, key);
 575         ret = -EBUSY;
 576
 577 bail:
 578         return ret;
 579 }
 580
 581 /**
 582  * ipath_manage_rcvq - manage a port's receive queue
 583  * @pd: the port
 584  * @start_stop: action to carry out
 585  *
 586  * start_stop == 0 disables receive on the port, for use in queue
 587  * overflow conditions.  start_stop==1 re-enables, to be used to
 588  * re-init the software copy of the head register
 589  */
 590 static int ipath_manage_rcvq(struct ipath_portdata *pd, int start_stop)
 591 {
 592         struct ipath_devdata *dd = pd->port_dd;
 593         u64 tval;
 594
 595         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u\n",
 596                    start_stop ? "en" : "dis", dd->ipath_unit,
 597                    pd->port_port);
 598         /* atomically clear receive enable port. */
 599         if (start_stop) {
 600                 /*
 601                  * On enable, force in-memory copy of the tail register to
 602                  * 0, so that protocol code doesn't have to worry about
 603                  * whether or not the chip has yet updated the in-memory
 604                  * copy or not on return from the system call. The chip
 605                  * always resets it's tail register back to 0 on a
 606                  * transition from disabled to enabled.  This could cause a
 607                  * problem if software was broken, and did the enable w/o
 608                  * the disable, but eventually the in-memory copy will be
 609                  * updated and correct itself, even in the face of software
 610                  * bugs.
 611                  */
 612                 *pd->port_rcvhdrtail_kvaddr = 0;
 613                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 614                         &dd->ipath_rcvctrl);
 615         } else
 616                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 617                           &dd->ipath_rcvctrl);
 618         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 619                          dd->ipath_rcvctrl);
 620         /* now be sure chip saw it before we return */
 621         tval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 622         if (start_stop) {
 623                 /*
 624                  * And try to be sure that tail reg update has happened too.
 625                  * This should in theory interlock with the RXE changes to
 626                  * the tail register.  Don't assign it to the tail register
 627                  * in memory copy, since we could overwrite an update by the
 628                  * chip if we did.
 629                  */
 630                 tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 631         }
 632         /* always; new head should be equal to new tail; see above */
 633         return 0;
 634 }
 635
 636 static void ipath_clean_part_key(struct ipath_portdata *pd,
 637                                  struct ipath_devdata *dd)
 638 {
 639         int i, j, pchanged = 0;
 640         u64 oldpkey;
 641
 642         /* for debugging only */
 643         oldpkey = (u64) dd->ipath_pkeys[0] |
 644                 ((u64) dd->ipath_pkeys[1] << 16) |
 645                 ((u64) dd->ipath_pkeys[2] << 32) |
 646                 ((u64) dd->ipath_pkeys[3] << 48);
 647
 648         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 649                 if (!pd->port_pkeys[i])
 650                         continue;
 651                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 652                            pd->port_pkeys[i]);
 653                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 654                         /* check for match independent of the global bit */
 655                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 656                             (pd->port_pkeys[i] & 0x7fff))
 657                                 continue;
 658                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 659                                 ipath_cdbg(VERBOSE, "p%u clear key "
 660                                            "%x matches #%d\n",
 661                                            pd->port_port,
 662                                            pd->port_pkeys[i], j);
 663                                 ipath_stats.sps_pkeys[j] =
 664                                         dd->ipath_pkeys[j] = 0;
 665                                 pchanged++;
 666                         }
 667                         else ipath_cdbg(
 668                                 VERBOSE, "p%u key %x matches #%d, "
 669                                 "but ref still %d\n", pd->port_port,
 670                                 pd->port_pkeys[i], j,
 671                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 672                         break;
 673                 }
 674                 pd->port_pkeys[i] = 0;
 675         }
 676         if (pchanged) {
 677                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 678                         ((u64) dd->ipath_pkeys[1] << 16) |
 679                         ((u64) dd->ipath_pkeys[2] << 32) |
 680                         ((u64) dd->ipath_pkeys[3] << 48);
 681                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 682                            "new pkey reg %llx\n", pd->port_port,
 683                            (unsigned long long) oldpkey,
 684                            (unsigned long long) pkey);
 685                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 686                                  pkey);
 687         }
 688 }
 689
 690 /**
 691  * ipath_create_user_egr - allocate eager TID buffers
 692  * @pd: the port to allocate TID buffers for
 693  *
 694  * This routine is now quite different for user and kernel, because
 695  * the kernel uses skb's, for the accelerated network performance
 696  * This is the user port version
 697  *
 698  * Allocate the eager TID buffers and program them into infinipath
 699  * They are no longer completely contiguous, we do multiple allocation
 700  * calls.
 701  */
 702 static int ipath_create_user_egr(struct ipath_portdata *pd)
 703 {
 704         struct ipath_devdata *dd = pd->port_dd;
 705         unsigned e, egrcnt, alloced, egrperchunk, chunk, egrsize, egroff;
 706         size_t size;
 707         int ret;
 708
 709         egrcnt = dd->ipath_rcvegrcnt;
 710         /* TID number offset for this port */
 711         egroff = pd->port_port * egrcnt;
 712         egrsize = dd->ipath_rcvegrbufsize;
 713         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 714                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 715
 716         /*
 717          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 718          * physically contiguous memory, advance through it until used up
 719          * and then allocate more.  Of course, we need memory to store those
 720          * extra pointers, now.  Started out with 256KB, but under heavy
 721          * memory pressure (creating large files and then copying them over
 722          * NFS while doing lots of MPI jobs), we hit some allocation
 723          * failures, even though we can sleep...  (2.6.10) Still get
 724          * failures at 64K.  32K is the lowest we can go without waiting
 725          * more memory again.  It seems likely that the coalescing in
 726          * free_pages, etc. still has issues (as it has had previously
 727          * during 2.6.x development).
 728          */
 729         size = 0x8000;
 730         alloced = ALIGN(egrsize * egrcnt, size);
 731         egrperchunk = size / egrsize;
 732         chunk = (egrcnt + egrperchunk - 1) / egrperchunk;
 733         pd->port_rcvegrbuf_chunks = chunk;
 734         pd->port_rcvegrbufs_perchunk = egrperchunk;
 735         pd->port_rcvegrbuf_size = size;
 736         pd->port_rcvegrbuf = vmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]));
 737         if (!pd->port_rcvegrbuf) {
 738                 ret = -ENOMEM;
 739                 goto bail;
 740         }
 741         pd->port_rcvegrbuf_phys =
 742                 vmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]));
 743         if (!pd->port_rcvegrbuf_phys) {
 744                 ret = -ENOMEM;
 745                 goto bail_rcvegrbuf;
 746         }
 747         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 748                 /*
 749                  * GFP_USER, but without GFP_FS, so buffer cache can be
 750                  * coalesced (we hope); otherwise, even at order 4,
 751                  * heavy filesystem activity makes these fail
 752                  */
 753                 gfp_t gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 754
 755                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 756                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 757                         gfp_flags);
 758
 759                 if (!pd->port_rcvegrbuf[e]) {
 760                         ret = -ENOMEM;
 761                         goto bail_rcvegrbuf_phys;
 762                 }
 763         }
 764
 765         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 766
 767         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 768                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 769                 unsigned i;
 770
 771                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 772                         dd->ipath_f_put_tid(dd, e + egroff +
 773                                             (u64 __iomem *)
 774                                             ((char __iomem *)
 775                                              dd->ipath_kregbase +
 776                                              dd->ipath_rcvegrbase), 0, pa);
 777                         pa += egrsize;
 778                 }
 779                 cond_resched(); /* don't hog the cpu */
 780         }
 781
 782         ret = 0;
 783         goto bail;
 784
 785 bail_rcvegrbuf_phys:
 786         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 787                      pd->port_rcvegrbuf[e]; e++)
 788                 dma_free_coherent(&dd->pcidev->dev, size,
 789                                   pd->port_rcvegrbuf[e],
 790                                   pd->port_rcvegrbuf_phys[e]);
 791
 792         vfree(pd->port_rcvegrbuf_phys);
 793         pd->port_rcvegrbuf_phys = NULL;
 794 bail_rcvegrbuf:
 795         vfree(pd->port_rcvegrbuf);
 796         pd->port_rcvegrbuf = NULL;
 797 bail:
 798         return ret;
 799 }
 800
 801 static int ipath_do_user_init(struct ipath_portdata *pd,
 802                               const struct ipath_user_info *uinfo)
 803 {
 804         int ret = 0;
 805         struct ipath_devdata *dd = pd->port_dd;
 806         u64 physaddr, uaddr, off, atmp;
 807         struct page *pagep;
 808         u32 head32;
 809         u64 head;
 810
 811         /* for now, if major version is different, bail */
 812         if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
 813                 dev_info(&dd->pcidev->dev,
 814                          "User major version %d not same as driver "
 815                          "major %d\n", uinfo->spu_userversion >> 16,
 816                          IPATH_USER_SWMAJOR);
 817                 ret = -ENODEV;
 818                 goto done;
 819         }
 820
 821         if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
 822                 ipath_dbg("User minor version %d not same as driver "
 823                           "minor %d\n", uinfo->spu_userversion & 0xffff,
 824                           IPATH_USER_SWMINOR);
 825
 826         if (uinfo->spu_rcvhdrsize) {
 827                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
 828                 if (ret)
 829                         goto done;
 830         }
 831
 832         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
 833
 834         /* set up for the rcvhdr Q tail register writeback to user memory */
 835         if (!uinfo->spu_rcvhdraddr ||
 836             !access_ok(VERIFY_WRITE, (u64 __user *) (unsigned long)
 837                        uinfo->spu_rcvhdraddr, sizeof(u64))) {
 838                 ipath_dbg("Port %d rcvhdrtail addr %llx not valid\n",
 839                           pd->port_port,
 840                           (unsigned long long) uinfo->spu_rcvhdraddr);
 841                 ret = -EINVAL;
 842                 goto done;
 843         }
 844
 845         off = offset_in_page(uinfo->spu_rcvhdraddr);
 846         uaddr = PAGE_MASK & (unsigned long) uinfo->spu_rcvhdraddr;
 847         ret = ipath_get_user_pages_nocopy(uaddr, &pagep);
 848         if (ret) {
 849                 dev_info(&dd->pcidev->dev, "Failed to lookup and lock "
 850                          "address %llx for rcvhdrtail: errno %d\n",
 851                          (unsigned long long) uinfo->spu_rcvhdraddr, -ret);
 852                 goto done;
 853         }
 854         ipath_stats.sps_pagelocks++;
 855         pd->port_rcvhdrtail_uaddr = uaddr;
 856         pd->port_rcvhdrtail_pagep = pagep;
 857         pd->port_rcvhdrtail_kvaddr =
 858                 page_address(pagep);
 859         pd->port_rcvhdrtail_kvaddr += off;
 860         physaddr = page_to_phys(pagep) + off;
 861         ipath_cdbg(VERBOSE, "port %d user addr %llx hdrtailaddr, %llx "
 862                    "physical (off=%llx)\n",
 863                    pd->port_port,
 864                    (unsigned long long) uinfo->spu_rcvhdraddr,
 865                    (unsigned long long) physaddr, (unsigned long long) off);
 866         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
 867                               pd->port_port, physaddr);
 868         atmp = ipath_read_kreg64_port(dd,
 869                                       dd->ipath_kregs->kr_rcvhdrtailaddr,
 870                                       pd->port_port);
 871         if (physaddr != atmp) {
 872                 ipath_dev_err(dd,
 873                               "Catastrophic software error, "
 874                               "RcvHdrTailAddr%u written as %llx, "
 875                               "read back as %llx\n", pd->port_port,
 876                               (unsigned long long) physaddr,
 877                               (unsigned long long) atmp);
 878                 ret = -EINVAL;
 879                 goto done;
 880         }
 881
 882         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
 883         pd->port_piobufs = dd->ipath_piobufbase +
 884                 dd->ipath_pbufsport * (pd->port_port -
 885                                        1) * dd->ipath_palign;
 886         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
 887                    pd->port_port, pd->port_piobufs);
 888
 889         /*
 890          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
 891          * array for time being.  If pd->port_port > chip-supported,
 892          * we need to do extra stuff here to handle by handling overflow
 893          * through port 0, someday
 894          */
 895         ret = ipath_create_rcvhdrq(dd, pd);
 896         if (!ret)
 897                 ret = ipath_create_user_egr(pd);
 898         if (ret)
 899                 goto done;
 900         /* enable receives now */
 901         /* atomically set enable bit for this port */
 902         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 903                 &dd->ipath_rcvctrl);
 904
 905         /*
 906          * set the head registers for this port to the current values
 907          * of the tail pointers, since we don't know if they were
 908          * updated on last use of the port.
 909          */
 910         head32 = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 911         head = (u64) head32;
 912         ipath_write_ureg(dd, ur_rcvhdrhead, head, pd->port_port);
 913         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
 914         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
 915         dd->ipath_lastegrheads[pd->port_port] = -1;
 916         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
 917         ipath_cdbg(VERBOSE, "Wrote port%d head %llx, egrhead %x from "
 918                    "tail regs\n", pd->port_port,
 919                    (unsigned long long) head, head32);
 920         pd->port_tidcursor = 0; /* start at beginning after open */
 921         /*
 922          * now enable the port; the tail registers will be written to memory
 923          * by the chip as soon as it sees the write to
 924          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
 925          * transition from 0 to 1, so clear it first, then set it as part of
 926          * enabling the port.  This will (very briefly) affect any other
 927          * open ports, but it shouldn't be long enough to be an issue.
 928          */
 929         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 930                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
 931         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 932                          dd->ipath_rcvctrl);
 933
 934 done:
 935         return ret;
 936 }
 937
 938 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
 939                      u64 ureg)
 940 {
 941         unsigned long phys;
 942         int ret;
 943
 944         /* it's the real hardware, so io_remap works */
 945
 946         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
 947                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
 948                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
 949                 ret = -EFAULT;
 950         } else {
 951                 phys = dd->ipath_physaddr + ureg;
 952                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 953
 954                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 955                 ret = io_remap_pfn_range(vma, vma->vm_start,
 956                                          phys >> PAGE_SHIFT,
 957                                          vma->vm_end - vma->vm_start,
 958                                          vma->vm_page_prot);
 959         }
 960         return ret;
 961 }
 962
 963 static int mmap_piobufs(struct vm_area_struct *vma,
 964                         struct ipath_devdata *dd,
 965                         struct ipath_portdata *pd)
 966 {
 967         unsigned long phys;
 968         int ret;
 969
 970         /*
 971          * When we map the PIO buffers, we want to map them as writeonly, no
 972          * read possible.
 973          */
 974
 975         if ((vma->vm_end - vma->vm_start) >
 976             (dd->ipath_pbufsport * dd->ipath_palign)) {
 977                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
 978                          "reqlen %lx > PAGE\n",
 979                          vma->vm_end - vma->vm_start);
 980                 ret = -EFAULT;
 981                 goto bail;
 982         }
 983
 984         phys = dd->ipath_physaddr + pd->port_piobufs;
 985         /*
 986          * Do *NOT* mark this as non-cached (PWT bit), or we don't get the
 987          * write combining behavior we want on the PIO buffers!
 988          * vma->vm_page_prot =
 989          *        pgprot_noncached(vma->vm_page_prot);
 990          */
 991
 992         if (vma->vm_flags & VM_READ) {
 993                 dev_info(&dd->pcidev->dev,
 994                          "Can't map piobufs as readable (flags=%lx)\n",
 995                          vma->vm_flags);
 996                 ret = -EPERM;
 997                 goto bail;
 998         }
 999
1000         /* don't allow them to later change to readable with mprotect */
1001
1002         vma->vm_flags &= ~VM_MAYWRITE;
1003         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1004
1005         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1006                                  vma->vm_end - vma->vm_start,
1007                                  vma->vm_page_prot);
1008 bail:
1009         return ret;
1010 }
1011
1012 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1013                            struct ipath_portdata *pd)
1014 {
1015         struct ipath_devdata *dd = pd->port_dd;
1016         unsigned long start, size;
1017         size_t total_size, i;
1018         dma_addr_t *phys;
1019         int ret;
1020
1021         if (!pd->port_rcvegrbuf) {
1022                 ret = -EFAULT;
1023                 goto bail;
1024         }
1025
1026         size = pd->port_rcvegrbuf_size;
1027         total_size = pd->port_rcvegrbuf_chunks * size;
1028         if ((vma->vm_end - vma->vm_start) > total_size) {
1029                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1030                          "reqlen %lx > actual %lx\n",
1031                          vma->vm_end - vma->vm_start,
1032                          (unsigned long) total_size);
1033                 ret = -EFAULT;
1034                 goto bail;
1035         }
1036
1037         if (vma->vm_flags & VM_WRITE) {
1038                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1039                          "writable (flags=%lx)\n", vma->vm_flags);
1040                 ret = -EPERM;
1041                 goto bail;
1042         }
1043
1044         start = vma->vm_start;
1045         phys = pd->port_rcvegrbuf_phys;
1046
1047         /* don't allow them to later change to writeable with mprotect */
1048         vma->vm_flags &= ~VM_MAYWRITE;
1049
1050         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1051                 ret = remap_pfn_range(vma, start, phys[i] >> PAGE_SHIFT,
1052                                       size, vma->vm_page_prot);
1053                 if (ret < 0)
1054                         goto bail;
1055         }
1056         ret = 0;
1057
1058 bail:
1059         return ret;
1060 }
1061
1062 static int mmap_rcvhdrq(struct vm_area_struct *vma,
1063                         struct ipath_portdata *pd)
1064 {
1065         struct ipath_devdata *dd = pd->port_dd;
1066         size_t total_size;
1067         int ret;
1068
1069         /*
1070          * kmalloc'ed memory, physically contiguous; this is from
1071          * spi_rcvhdr_base; we allow user to map read-write so they can
1072          * write hdrq entries to allow protocol code to directly poll
1073          * whether a hdrq entry has been written.
1074          */
1075         total_size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1076                            sizeof(u32), PAGE_SIZE);
1077         if ((vma->vm_end - vma->vm_start) > total_size) {
1078                 dev_info(&dd->pcidev->dev,
1079                          "FAIL on rcvhdrq: reqlen %lx > actual %lx\n",
1080                          vma->vm_end - vma->vm_start,
1081                          (unsigned long) total_size);
1082                 ret = -EFAULT;
1083                 goto bail;
1084         }
1085
1086         ret = remap_pfn_range(vma, vma->vm_start,
1087                               pd->port_rcvhdrq_phys >> PAGE_SHIFT,
1088                               vma->vm_end - vma->vm_start,
1089                               vma->vm_page_prot);
1090 bail:
1091         return ret;
1092 }
1093
1094 static int mmap_pioavailregs(struct vm_area_struct *vma,
1095                              struct ipath_portdata *pd)
1096 {
1097         struct ipath_devdata *dd = pd->port_dd;
1098         int ret;
1099
1100         /*
1101          * when we map the PIO bufferavail registers, we want to map them as
1102          * readonly, no write possible.
1103          *
1104          * kmalloc'ed memory, physically contiguous, one page only, readonly
1105          */
1106
1107         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1108                 dev_info(&dd->pcidev->dev, "FAIL on pioavailregs_dma: "
1109                          "reqlen %lx > actual %lx\n",
1110                          vma->vm_end - vma->vm_start,
1111                          (unsigned long) PAGE_SIZE);
1112                 ret = -EFAULT;
1113                 goto bail;
1114         }
1115
1116         if (vma->vm_flags & VM_WRITE) {
1117                 dev_info(&dd->pcidev->dev,
1118                          "Can't map pioavailregs as writable (flags=%lx)\n",
1119                          vma->vm_flags);
1120                 ret = -EPERM;
1121                 goto bail;
1122         }
1123
1124         /* don't allow them to later change with mprotect */
1125         vma->vm_flags &= ~VM_MAYWRITE;
1126
1127         ret = remap_pfn_range(vma, vma->vm_start,
1128                               dd->ipath_pioavailregs_phys >> PAGE_SHIFT,
1129                               PAGE_SIZE, vma->vm_page_prot);
1130 bail:
1131         return ret;
1132 }
1133
1134 /**
1135  * ipath_mmap - mmap various structures into user space
1136  * @fp: the file pointer
1137  * @vma: the VM area
1138  *
1139  * We use this to have a shared buffer between the kernel and the user code
1140  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1141  * buffers in the chip.  We have the open and close entries so we can bump
1142  * the ref count and keep the driver from being unloaded while still mapped.
1143  */
1144 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1145 {
1146         struct ipath_portdata *pd;
1147         struct ipath_devdata *dd;
1148         u64 pgaddr, ureg;
1149         int ret;
1150
1151         pd = port_fp(fp);
1152         dd = pd->port_dd;
1153         /*
1154          * This is the ipath_do_user_init() code, mapping the shared buffers
1155          * into the user process. The address referred to by vm_pgoff is the
1156          * virtual, not physical, address; we only do one mmap for each
1157          * space mapped.
1158          */
1159         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1160
1161         /*
1162          * note that ureg does *NOT* have the kregvirt as part of it, to be
1163          * sure that for 32 bit programs, we don't end up trying to map a >
1164          * 44 address.  Has to match ipath_get_base_info() code that sets
1165          * __spi_uregbase
1166          */
1167
1168         ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1169
1170         ipath_cdbg(MM, "ushare: pgaddr %llx vm_start=%lx, vmlen %lx\n",
1171                    (unsigned long long) pgaddr, vma->vm_start,
1172                    vma->vm_end - vma->vm_start);
1173
1174         if (pgaddr == ureg)
1175                 ret = mmap_ureg(vma, dd, ureg);
1176         else if (pgaddr == pd->port_piobufs)
1177                 ret = mmap_piobufs(vma, dd, pd);
1178         else if (pgaddr == (u64) pd->port_rcvegr_phys)
1179                 ret = mmap_rcvegrbufs(vma, pd);
1180         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1181                 ret = mmap_rcvhdrq(vma, pd);
1182         else if (pgaddr == dd->ipath_pioavailregs_phys)
1183                 ret = mmap_pioavailregs(vma, pd);
1184         else
1185                 ret = -EINVAL;
1186
1187         vma->vm_private_data = NULL;
1188
1189         if (ret < 0)
1190                 dev_info(&dd->pcidev->dev,
1191                          "Failure %d on addr %lx, off %lx\n",
1192                          -ret, vma->vm_start, vma->vm_pgoff);
1193
1194         return ret;
1195 }
1196
1197 static unsigned int ipath_poll(struct file *fp,
1198                                struct poll_table_struct *pt)
1199 {
1200         struct ipath_portdata *pd;
1201         u32 head, tail;
1202         int bit;
1203         struct ipath_devdata *dd;
1204
1205         pd = port_fp(fp);
1206         dd = pd->port_dd;
1207
1208         bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1209         set_bit(bit, &dd->ipath_rcvctrl);
1210
1211         /*
1212          * Before blocking, make sure that head is still == tail,
1213          * reading from the chip, so we can be sure the interrupt
1214          * enable has made it to the chip.  If not equal, disable
1215          * interrupt again and return immediately.  This avoids races,
1216          * and the overhead of the chip read doesn't matter much at
1217          * this point, since we are waiting for something anyway.
1218          */
1219
1220         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1221                          dd->ipath_rcvctrl);
1222
1223         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1224         tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1225
1226         if (tail == head) {
1227                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1228                 if(dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1229                         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1230                                                dd->ipath_rhdrhead_intr_off
1231                                                | head, pd->port_port);
1232                 poll_wait(fp, &pd->port_wait, pt);
1233
1234                 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1235                         /* timed out, no packets received */
1236                         clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1237                         pd->port_rcvwait_to++;
1238                 }
1239         }
1240         else {
1241                 /* it's already happened; don't do wait_event overhead */
1242                 pd->port_rcvnowait++;
1243         }
1244
1245         clear_bit(bit, &dd->ipath_rcvctrl);
1246         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1247                          dd->ipath_rcvctrl);
1248
1249         return 0;
1250 }
1251
1252 static int try_alloc_port(struct ipath_devdata *dd, int port,
1253                           struct file *fp)
1254 {
1255         int ret;
1256
1257         if (!dd->ipath_pd[port]) {
1258                 void *p, *ptmp;
1259
1260                 p = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1261
1262                 /*
1263                  * Allocate memory for use in ipath_tid_update() just once
1264                  * at open, not per call.  Reduces cost of expected send
1265                  * setup.
1266                  */
1267                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1268                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1269                                GFP_KERNEL);
1270                 if (!p || !ptmp) {
1271                         ipath_dev_err(dd, "Unable to allocate portdata "
1272                                       "memory, failing open\n");
1273                         ret = -ENOMEM;
1274                         kfree(p);
1275                         kfree(ptmp);
1276                         goto bail;
1277                 }
1278                 dd->ipath_pd[port] = p;
1279                 dd->ipath_pd[port]->port_port = port;
1280                 dd->ipath_pd[port]->port_dd = dd;
1281                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1282                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1283         }
1284         if (!dd->ipath_pd[port]->port_cnt) {
1285                 dd->ipath_pd[port]->port_cnt = 1;
1286                 fp->private_data = (void *) dd->ipath_pd[port];
1287                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1288                            current->comm, current->pid, dd->ipath_unit,
1289                            port);
1290                 dd->ipath_pd[port]->port_pid = current->pid;
1291                 strncpy(dd->ipath_pd[port]->port_comm, current->comm,
1292                         sizeof(dd->ipath_pd[port]->port_comm));
1293                 ipath_stats.sps_ports++;
1294                 ret = 0;
1295                 goto bail;
1296         }
1297         ret = -EBUSY;
1298
1299 bail:
1300         return ret;
1301 }
1302
1303 static inline int usable(struct ipath_devdata *dd)
1304 {
1305         return dd &&
1306                 (dd->ipath_flags & IPATH_PRESENT) &&
1307                 dd->ipath_kregbase &&
1308                 dd->ipath_lid &&
1309                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1310                                      | IPATH_LINKUNK));
1311 }
1312
1313 static int find_free_port(int unit, struct file *fp)
1314 {
1315         struct ipath_devdata *dd = ipath_lookup(unit);
1316         int ret, i;
1317
1318         if (!dd) {
1319                 ret = -ENODEV;
1320                 goto bail;
1321         }
1322
1323         if (!usable(dd)) {
1324                 ret = -ENETDOWN;
1325                 goto bail;
1326         }
1327
1328         for (i = 0; i < dd->ipath_cfgports; i++) {
1329                 ret = try_alloc_port(dd, i, fp);
1330                 if (ret != -EBUSY)
1331                         goto bail;
1332         }
1333         ret = -EBUSY;
1334
1335 bail:
1336         return ret;
1337 }
1338
1339 static int find_best_unit(struct file *fp)
1340 {
1341         int ret = 0, i, prefunit = -1, devmax;
1342         int maxofallports, npresent, nup;
1343         int ndev;
1344
1345         (void) ipath_count_units(&npresent, &nup, &maxofallports);
1346
1347         /*
1348          * This code is present to allow a knowledgeable person to
1349          * specify the layout of processes to processors before opening
1350          * this driver, and then we'll assign the process to the "closest"
1351          * HT-400 to that processor (we assume reasonable connectivity,
1352          * for now).  This code assumes that if affinity has been set
1353          * before this point, that at most one cpu is set; for now this
1354          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1355          * in case some kernel variant sets none of the bits when no
1356          * affinity is set.  2.6.11 and 12 kernels have all present
1357          * cpus set.  Some day we'll have to fix it up further to handle
1358          * a cpu subset.  This algorithm fails for two HT-400's connected
1359          * in tunnel fashion.  Eventually this needs real topology
1360          * information.  There may be some issues with dual core numbering
1361          * as well.  This needs more work prior to release.
1362          */
1363         if (!cpus_empty(current->cpus_allowed) &&
1364             !cpus_full(current->cpus_allowed)) {
1365                 int ncpus = num_online_cpus(), curcpu = -1;
1366                 for (i = 0; i < ncpus; i++)
1367                         if (cpu_isset(i, current->cpus_allowed)) {
1368                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1369                                            "cpu %d\n", current->comm,
1370                                            current->pid, i);
1371                                 curcpu = i;
1372                         }
1373                 if (curcpu != -1) {
1374                         if (npresent) {
1375                                 prefunit = curcpu / (ncpus / npresent);
1376                                 ipath_dbg("%s[%u] %d chips, %d cpus, "
1377                                           "%d cpus/chip, select unit %d\n",
1378                                           current->comm, current->pid,
1379                                           npresent, ncpus, ncpus / npresent,
1380                                           prefunit);
1381                         }
1382                 }
1383         }
1384
1385         /*
1386          * user ports start at 1, kernel port is 0
1387          * For now, we do round-robin access across all chips
1388          */
1389
1390         if (prefunit != -1)
1391                 devmax = prefunit + 1;
1392         else
1393                 devmax = ipath_count_units(NULL, NULL, NULL);
1394 recheck:
1395         for (i = 1; i < maxofallports; i++) {
1396                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1397                      ndev++) {
1398                         struct ipath_devdata *dd = ipath_lookup(ndev);
1399
1400                         if (!usable(dd))
1401                                 continue; /* can't use this unit */
1402                         if (i >= dd->ipath_cfgports)
1403                                 /*
1404                                  * Maxed out on users of this unit. Try
1405                                  * next.
1406                                  */
1407                                 continue;
1408                         ret = try_alloc_port(dd, i, fp);
1409                         if (!ret)
1410                                 goto done;
1411                 }
1412         }
1413
1414         if (npresent) {
1415                 if (nup == 0) {
1416                         ret = -ENETDOWN;
1417                         ipath_dbg("No ports available (none initialized "
1418                                   "and ready)\n");
1419                 } else {
1420                         if (prefunit > 0) {
1421                                 /* if started above 0, retry from 0 */
1422                                 ipath_cdbg(PROC,
1423                                            "%s[%u] no ports on prefunit "
1424                                            "%d, clear and re-check\n",
1425                                            current->comm, current->pid,
1426                                            prefunit);
1427                                 devmax = ipath_count_units(NULL, NULL,
1428                                                            NULL);
1429                                 prefunit = -1;
1430                                 goto recheck;
1431                         }
1432                         ret = -EBUSY;
1433                         ipath_dbg("No ports available\n");
1434                 }
1435         } else {
1436                 ret = -ENXIO;
1437                 ipath_dbg("No boards found\n");
1438         }
1439
1440 done:
1441         return ret;
1442 }
1443
1444 static int ipath_open(struct inode *in, struct file *fp)
1445 {
1446         int ret, user_minor;
1447
1448         mutex_lock(&ipath_mutex);
1449
1450         user_minor = iminor(in) - IPATH_USER_MINOR_BASE;
1451         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1452                    (long)in->i_rdev, user_minor);
1453
1454         if (user_minor)
1455                 ret = find_free_port(user_minor - 1, fp);
1456         else
1457                 ret = find_best_unit(fp);
1458
1459         mutex_unlock(&ipath_mutex);
1460         return ret;
1461 }
1462
1463 /**
1464  * unlock_exptid - unlock any expected TID entries port still had in use
1465  * @pd: port
1466  *
1467  * We don't actually update the chip here, because we do a bulk update
1468  * below, using ipath_f_clear_tids.
1469  */
1470 static void unlock_expected_tids(struct ipath_portdata *pd)
1471 {
1472         struct ipath_devdata *dd = pd->port_dd;
1473         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1474         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1475
1476         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1477                    pd->port_port);
1478         for (i = port_tidbase; i < maxtid; i++) {
1479                 if (!dd->ipath_pageshadow[i])
1480                         continue;
1481
1482                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1483                                                   1);
1484                 dd->ipath_pageshadow[i] = NULL;
1485                 cnt++;
1486                 ipath_stats.sps_pageunlocks++;
1487         }
1488         if (cnt)
1489                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1490                            pd->port_port, cnt);
1491
1492         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1493                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1494                            (unsigned long long) ipath_stats.sps_pagelocks,
1495                            (unsigned long long)
1496                            ipath_stats.sps_pageunlocks);
1497 }
1498
1499 static int ipath_close(struct inode *in, struct file *fp)
1500 {
1501         int ret = 0;
1502         struct ipath_portdata *pd;
1503         struct ipath_devdata *dd;
1504         unsigned port;
1505
1506         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1507                    (long)in->i_rdev, fp->private_data);
1508
1509         mutex_lock(&ipath_mutex);
1510
1511         pd = port_fp(fp);
1512         port = pd->port_port;
1513         fp->private_data = NULL;
1514         dd = pd->port_dd;
1515
1516         if (pd->port_hdrqfull) {
1517                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1518                            "during run\n", pd->port_comm, pd->port_pid,
1519                            pd->port_hdrqfull);
1520                 pd->port_hdrqfull = 0;
1521         }
1522
1523         if (pd->port_rcvwait_to || pd->port_piowait_to
1524             || pd->port_rcvnowait || pd->port_pionowait) {
1525                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1526                            "%u rcv %u, pio already\n",
1527                            pd->port_port, pd->port_rcvwait_to,
1528                            pd->port_piowait_to, pd->port_rcvnowait,
1529                            pd->port_pionowait);
1530                 pd->port_rcvwait_to = pd->port_piowait_to =
1531                         pd->port_rcvnowait = pd->port_pionowait = 0;
1532         }
1533         if (pd->port_flag) {
1534                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
1535                           pd->port_port, pd->port_flag);
1536                 pd->port_flag = 0;
1537         }
1538
1539         if (dd->ipath_kregbase) {
1540                 if (pd->port_rcvhdrtail_uaddr) {
1541                         pd->port_rcvhdrtail_uaddr = 0;
1542                         pd->port_rcvhdrtail_kvaddr = NULL;
1543                         ipath_release_user_pages_on_close(
1544                                 &pd->port_rcvhdrtail_pagep, 1);
1545                         pd->port_rcvhdrtail_pagep = NULL;
1546                         ipath_stats.sps_pageunlocks++;
1547                 }
1548                 ipath_write_kreg_port(
1549                         dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1550                         port, 0ULL);
1551                 ipath_write_kreg_port(
1552                         dd, dd->ipath_kregs->kr_rcvhdraddr,
1553                         pd->port_port, 0);
1554
1555                 /* clean up the pkeys for this port user */
1556                 ipath_clean_part_key(pd, dd);
1557
1558                 if (port < dd->ipath_cfgports) {
1559                         int i = dd->ipath_pbufsport * (port - 1);
1560                         ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
1561
1562                         /* atomically clear receive enable port. */
1563                         clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
1564                                   &dd->ipath_rcvctrl);
1565                         ipath_write_kreg(
1566                                 dd,
1567                                 dd->ipath_kregs->kr_rcvctrl,
1568                                 dd->ipath_rcvctrl);
1569
1570                         if (dd->ipath_pageshadow)
1571                                 unlock_expected_tids(pd);
1572                         ipath_stats.sps_ports--;
1573                         ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
1574                                    pd->port_comm, pd->port_pid,
1575                                    dd->ipath_unit, port);
1576                 }
1577         }
1578
1579         pd->port_cnt = 0;
1580         pd->port_pid = 0;
1581
1582         dd->ipath_f_clear_tids(dd, pd->port_port);
1583
1584         ipath_free_pddata(dd, pd->port_port, 0);
1585
1586         mutex_unlock(&ipath_mutex);
1587
1588         return ret;
1589 }
1590
1591 static int ipath_port_info(struct ipath_portdata *pd,
1592                            struct ipath_port_info __user *uinfo)
1593 {
1594         struct ipath_port_info info;
1595         int nup;
1596         int ret;
1597
1598         (void) ipath_count_units(NULL, &nup, NULL);
1599         info.num_active = nup;
1600         info.unit = pd->port_dd->ipath_unit;
1601         info.port = pd->port_port;
1602
1603         if (copy_to_user(uinfo, &info, sizeof(info))) {
1604                 ret = -EFAULT;
1605                 goto bail;
1606         }
1607         ret = 0;
1608
1609 bail:
1610         return ret;
1611 }
1612
1613 static ssize_t ipath_write(struct file *fp, const char __user *data,
1614                            size_t count, loff_t *off)
1615 {
1616         const struct ipath_cmd __user *ucmd;
1617         struct ipath_portdata *pd;
1618         const void __user *src;
1619         size_t consumed, copy;
1620         struct ipath_cmd cmd;
1621         ssize_t ret = 0;
1622         void *dest;
1623
1624         if (count < sizeof(cmd.type)) {
1625                 ret = -EINVAL;
1626                 goto bail;
1627         }
1628
1629         ucmd = (const struct ipath_cmd __user *) data;
1630
1631         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
1632                 ret = -EFAULT;
1633                 goto bail;
1634         }
1635
1636         consumed = sizeof(cmd.type);
1637
1638         switch (cmd.type) {
1639         case IPATH_CMD_USER_INIT:
1640                 copy = sizeof(cmd.cmd.user_info);
1641                 dest = &cmd.cmd.user_info;
1642                 src = &ucmd->cmd.user_info;
1643                 break;
1644         case IPATH_CMD_RECV_CTRL:
1645                 copy = sizeof(cmd.cmd.recv_ctrl);
1646                 dest = &cmd.cmd.recv_ctrl;
1647                 src = &ucmd->cmd.recv_ctrl;
1648                 break;
1649         case IPATH_CMD_PORT_INFO:
1650                 copy = sizeof(cmd.cmd.port_info);
1651                 dest = &cmd.cmd.port_info;
1652                 src = &ucmd->cmd.port_info;
1653                 break;
1654         case IPATH_CMD_TID_UPDATE:
1655         case IPATH_CMD_TID_FREE:
1656                 copy = sizeof(cmd.cmd.tid_info);
1657                 dest = &cmd.cmd.tid_info;
1658                 src = &ucmd->cmd.tid_info;
1659                 break;
1660         case IPATH_CMD_SET_PART_KEY:
1661                 copy = sizeof(cmd.cmd.part_key);
1662                 dest = &cmd.cmd.part_key;
1663                 src = &ucmd->cmd.part_key;
1664                 break;
1665         default:
1666                 ret = -EINVAL;
1667                 goto bail;
1668         }
1669
1670         if ((count - consumed) < copy) {
1671                 ret = -EINVAL;
1672                 goto bail;
1673         }
1674
1675         if (copy_from_user(dest, src, copy)) {
1676                 ret = -EFAULT;
1677                 goto bail;
1678         }
1679
1680         consumed += copy;
1681         pd = port_fp(fp);
1682
1683         switch (cmd.type) {
1684         case IPATH_CMD_USER_INIT:
1685                 ret = ipath_do_user_init(pd, &cmd.cmd.user_info);
1686                 if (ret < 0)
1687                         goto bail;
1688                 ret = ipath_get_base_info(
1689                         pd, (void __user *) (unsigned long)
1690                         cmd.cmd.user_info.spu_base_info,
1691                         cmd.cmd.user_info.spu_base_info_size);
1692                 break;
1693         case IPATH_CMD_RECV_CTRL:
1694                 ret = ipath_manage_rcvq(pd, cmd.cmd.recv_ctrl);
1695                 break;
1696         case IPATH_CMD_PORT_INFO:
1697                 ret = ipath_port_info(pd,
1698                                       (struct ipath_port_info __user *)
1699                                       (unsigned long) cmd.cmd.port_info);
1700                 break;
1701         case IPATH_CMD_TID_UPDATE:
1702                 ret = ipath_tid_update(pd, &cmd.cmd.tid_info);
1703                 break;
1704         case IPATH_CMD_TID_FREE:
1705                 ret = ipath_tid_free(pd, &cmd.cmd.tid_info);
1706                 break;
1707         case IPATH_CMD_SET_PART_KEY:
1708                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
1709                 break;
1710         }
1711
1712         if (ret >= 0)
1713                 ret = consumed;
1714
1715 bail:
1716         return ret;
1717 }
1718
1719 static struct class *ipath_class;
1720
1721 static int init_cdev(int minor, char *name, struct file_operations *fops,
1722                      struct cdev **cdevp, struct class_device **class_devp)
1723 {
1724         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
1725         struct cdev *cdev = NULL;
1726         struct class_device *class_dev = NULL;
1727         int ret;
1728
1729         cdev = cdev_alloc();
1730         if (!cdev) {
1731                 printk(KERN_ERR IPATH_DRV_NAME
1732                        ": Could not allocate cdev for minor %d, %s\n",
1733                        minor, name);
1734                 ret = -ENOMEM;
1735                 goto done;
1736         }
1737
1738         cdev->owner = THIS_MODULE;
1739         cdev->ops = fops;
1740         kobject_set_name(&cdev->kobj, name);
1741
1742         ret = cdev_add(cdev, dev, 1);
1743         if (ret < 0) {
1744                 printk(KERN_ERR IPATH_DRV_NAME
1745                        ": Could not add cdev for minor %d, %s (err %d)\n",
1746                        minor, name, -ret);
1747                 goto err_cdev;
1748         }
1749
1750         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
1751
1752         if (IS_ERR(class_dev)) {
1753                 ret = PTR_ERR(class_dev);
1754                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
1755                        "class_dev for minor %d, %s (err %d)\n",
1756                        minor, name, -ret);
1757                 goto err_cdev;
1758         }
1759
1760         goto done;
1761
1762 err_cdev:
1763         cdev_del(cdev);
1764         cdev = NULL;
1765
1766 done:
1767         if (ret >= 0) {
1768                 *cdevp = cdev;
1769                 *class_devp = class_dev;
1770         } else {
1771                 *cdevp = NULL;
1772                 *class_devp = NULL;
1773         }
1774
1775         return ret;
1776 }
1777
1778 int ipath_cdev_init(int minor, char *name, struct file_operations *fops,
1779                     struct cdev **cdevp, struct class_device **class_devp)
1780 {
1781         return init_cdev(minor, name, fops, cdevp, class_devp);
1782 }
1783
1784 static void cleanup_cdev(struct cdev **cdevp,
1785                          struct class_device **class_devp)
1786 {
1787         struct class_device *class_dev = *class_devp;
1788
1789         if (class_dev) {
1790                 class_device_unregister(class_dev);
1791                 *class_devp = NULL;
1792         }
1793
1794         if (*cdevp) {
1795                 cdev_del(*cdevp);
1796                 *cdevp = NULL;
1797         }
1798 }
1799
1800 void ipath_cdev_cleanup(struct cdev **cdevp,
1801                         struct class_device **class_devp)
1802 {
1803         cleanup_cdev(cdevp, class_devp);
1804 }
1805
1806 static struct cdev *wildcard_cdev;
1807 static struct class_device *wildcard_class_dev;
1808
1809 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
1810
1811 static int user_init(void)
1812 {
1813         int ret;
1814
1815         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
1816         if (ret < 0) {
1817                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
1818                        "chrdev region (err %d)\n", -ret);
1819                 goto done;
1820         }
1821
1822         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
1823
1824         if (IS_ERR(ipath_class)) {
1825                 ret = PTR_ERR(ipath_class);
1826                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
1827                        "device class (err %d)\n", -ret);
1828                 goto bail;
1829         }
1830
1831         goto done;
1832 bail:
1833         unregister_chrdev_region(dev, IPATH_NMINORS);
1834 done:
1835         return ret;
1836 }
1837
1838 static void user_cleanup(void)
1839 {
1840         if (ipath_class) {
1841                 class_destroy(ipath_class);
1842                 ipath_class = NULL;
1843         }
1844
1845         unregister_chrdev_region(dev, IPATH_NMINORS);
1846 }
1847
1848 static atomic_t user_count = ATOMIC_INIT(0);
1849 static atomic_t user_setup = ATOMIC_INIT(0);
1850
1851 int ipath_user_add(struct ipath_devdata *dd)
1852 {
1853         char name[10];
1854         int ret;
1855
1856         if (atomic_inc_return(&user_count) == 1) {
1857                 ret = user_init();
1858                 if (ret < 0) {
1859                         ipath_dev_err(dd, "Unable to set up user support: "
1860                                       "error %d\n", -ret);
1861                         goto bail;
1862                 }
1863                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
1864                                 &wildcard_class_dev);
1865                 if (ret < 0) {
1866                         ipath_dev_err(dd, "Could not create wildcard "
1867                                       "minor: error %d\n", -ret);
1868                         goto bail_sma;
1869                 }
1870
1871                 atomic_set(&user_setup, 1);
1872         }
1873
1874         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
1875
1876         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
1877                         &dd->user_cdev, &dd->user_class_dev);
1878         if (ret < 0)
1879                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
1880                               dd->ipath_unit + 1, name);
1881
1882         goto bail;
1883
1884 bail_sma:
1885         user_cleanup();
1886 bail:
1887         return ret;
1888 }
1889
1890 void ipath_user_remove(struct ipath_devdata *dd)
1891 {
1892         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
1893
1894         if (atomic_dec_return(&user_count) == 0) {
1895                 if (atomic_read(&user_setup) == 0)
1896                         goto bail;
1897
1898                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
1899                 user_cleanup();
1900
1901                 atomic_set(&user_setup, 0);
1902         }
1903 bail:
1904         return;
1905 }